Science of AI
What We Do
AI is advancing at an unprecedented pace, leaving significant gaps in our understanding of the fundamental principles fueling its success. The increasing training complexity and size of AI models pose significant challenges for scientists who study the underlying principles at work in ML models. We are working to develop and test new scientific, engineering, and mathematical principles underpinning the numerous technological breakthroughs in AI and deep learning.
Representational Learning
A New Geometric Theory of Representation in AI Vision Models
Research fellows Thomas Fel and Binxu Wang led a team that has developed a new way to interpret representations of visual information in DINOv2, a Vision Transformer model. The team proposes a new geometric perspective known as the Minkowski Representation Hypothesis, which claims that visual representations are organized in a manner that goes beyond linear sparsity.
Research Projects
In-Context Learning
Solvable Model of In-Context Learning Using Linear Attention
Kempner researchers provide a novel characterization of in-context learning (ICL) in an analytically solvable model, which offers insights into the sample complexity and data quality requirements for ICL. These insights can be applied to more complex, realistic architectures.
Diffusion Models
The Hidden Linear Structure in Diffusion Models and its Application in Analytical Teleportation
Kempner researchers analyze how diffusion models iteratively denoise white noise into structured data via learned score functions. The team uses theory and experiments to demonstrate that these score functions are dominated by a linear Gaussian component.
Sparse Autoencoders (SAEs)
What Data Assumptions Come With Your SAE?
A Kempner team shows that SAEs are inherently biased to detect only a subset of concepts in model activations shaped by their internal assumptions, highlighting the need to design SAE architectures in a manner that is aware of concept geometry.
